Research method
A 100-company library, built from visible evidence.
100 direct website studies · 0 documented fallbacks · source refreshed 2026-09-03Select
Take active companies with public websites from the YC directory mirror, sort by directory launch timestamp, then freeze the newest 100 slugs in the repository.
Capture
Render each public homepage at 1600 × 1000, scan up to 30,000 px, and preserve hero, middle, closing, and 390 px mobile evidence when those regions exist. A fingerprint ties each study to its committed screens.
Observe
Measure semantic color roles, type scale, spacing, radii, borders, shadows, components, section order, imagery, motion declarations, and responsive changes from the rendered page.
Translate
For direct captures, generate the human study, DESIGN.md, CSS variables, Tailwind v4 theme, and tokens JSON from one versioned record. All 100 companies retain raw observation JSON; unavailable captures receive no inferred design artifacts.
Publication rules
- No invented observations
- If a page cannot be rendered, its analysis stays unavailable.
- Timestamp every capture
- Websites change. Each study describes a particular observed version.
- Keep source and interpretation separate
- Company descriptions come from YC; interface concepts come from the captured homepage.
- Commit the evidence
- Metadata, deep analysis, screen evidence, and every agent export live with the application code.
- Mark review state
- Automated measurements remain clearly separate from studies that receive an editorial review.
- Disclose partial coverage
- Missing font faces and uninspectable embedded components are named in each study’s capture notes. Component previews demonstrate observed tokens; they are not original source components.
- State the integrity boundary
- Capture assumes a normal public site, not one deliberately falsifying browser DOM or style APIs.